the 5th of June 2013
PhD in Economics
Emmanuelle LAVAINE
Social inequalities linked to the effect of pollution on health
Thesis director
Mireille Chiroleu Assouline - Professor at Paris 1 Panthéon-Sorbonne University Carmen Arguedas Tomás - Professor at Autonomous University of Madrid
Introduction 2
1 Atmospheric Pollution, environmental disparities and Mortality Rate: An Econo-
metric Analysis 4
1.1 Introduction . . . . 5
1.2 Medical perspective . . . . 10
1.3 Presentation of the dataset . . . . 13
1.4 Model and Econometrics . . . . 23
1.4.1 Specification . . . . 23
1.4.2 Results . . . . 29
1.5 Conclusion . . . . 45
1.6 Appendix . . . . 47
2 Energy Production and Health Externalities: Evidence from Oil Strike Refiner- ies in France.1 49 2.1 Introduction . . . . 50
2.2 Background: Refineries, Air pollution and Health . . . . 55
2.2.1 Pollution and the refinery closure . . . . 55 1. This chapter is a joint work with Matthew Neidell from Columbia University
iii
2.2.2 Pollution and health . . . . 56
2.3 Data and empirical strategy . . . . 58
2.3.1 Data sources . . . . 58
2.3.2 Merging data . . . . 65
2.3.3 Empirical Methodology . . . . 66
2.4 Results . . . . 67
2.4.1 Refinery closures and pollution levels . . . . 67
2.4.2 Refinery closures and birth outcomes . . . . 70
2.5 Conclusion . . . . 77
3 The price of pollution and health: an hedonic approach 79 3.1 Introduction . . . . 80
3.2 Pollution, health and refinery closure . . . . 87
3.2.1 SO2pollution and health . . . . 87
3.2.2 Pollution and the refinery closure . . . . 88
3.2.3 The incidence of the refinery closure on housing prices . . . . 90
3.3 Dataset presentation . . . . 92
3.3.1 Pollution data . . . . 92
3.3.2 Morbidity data . . . . 95
3.3.3 Weather data and socioeconomic data . . . . 95
3.3.4 Property Prices . . . . 97
3.4 Estimation . . . . 97
3.5 Results . . . . 100
3.5.1 Pollution concentration and refinery closure . . . . 100
3.5.2 Respiratory outcomes and refinery closure . . . . 103
3.5.3 Property prices and refinery closure . . . . 108
3.5.4 Monetary evaluation . . . . 124 3.6 Conclusion . . . . 126
General conclusions 128
Bibliography 131
Many pollutants are declining throughout the industrialized world. However, exposure to air pollution, even at the levels commonly achieved nowadays in European countries, still leads to adverse health effects. In this context, there has been increasing global concern over the public health impacts attributed to environmental pollution.
We propose to investigate the causal effect of air pollution on infants’ health, respiratory outcomes in France using several natural experiments and a unique dataset combining data on environmental quality, health and property prices. Recently, the role of clean air policies have been increasing along with the rise of public health concerns in Europe. Given this increasing needs worldwide, further studies of air pollution policies are useful in order to better inform this important public policy process.
The first objective of the thesis is to explore empirically the relations between socio-economic status, environmental exposures and health outcomes. We also went further in the analysis of social inequalities linked to environmental pollution by sheding light on their macroeconomic con- sequences. I study differences in exposition and underline their consequences on morbidity by developing a small area empirical approach. First, I empirically measure the impact of pollution on health and productivity and look at how pollution can contribute to health inequalities. To do so, I estimate the relationship in France between nitrogen dioxyde (NO2), environmental disparities
2
and non incidental mortality rates. This study is part of new research on environmental justice, and provides an overview of the distribution of environmental risks. Second, we estimate the health ex- ternalities from oil production by exploiting the oil refinery strikes in France in October 2010. The strikes provide a natural experiment that enables us to overcome the typical omitted variable bias that arises from Tiebout sorting (Banzhaf and Walsh, 2008) (Greenstone 2003). Amid nationwide protests over pensions reform and broader concerns about oil industry practices in France, striking workers blocked refineries, which resulted in a complete brought cessation of operations to a halt at several major refineries for nearly a month. As we demonstrate, this lead to a sharp reduction in SO2 that which quickly dissipated rose again once the strike was resolved and production resumed.
We exploit this temporal event to identify the infant health externalities at birth from oil produc- tion, comparing outcomes in areas close to the refineries before and after the strike vs. during the strike, using areas far from the refineries as a control group. Finally, I have been developing a third research project focusing on the link between respiratory outcomes, housing prices and pollution using the hedonic price method. I try to draw inferences about individuals’ valuations of risk by combining estimates of the effect of air pollution on both property values and hospital respiratory admissions for respiratory causes in France. The analysis focuses on Dunkerque, a french census track in the Nord-Pas-de-Calais region in France where residents have recently experienced a re- finery closure. Housing prices are compared before and after the closure with the nearby census tracks within 50 kilometers, acting as a control group.
Atmospheric Pollution,
environmental disparities and
Mortality Rate: An Econometric Analysis
4
This paper presents the first study of environmental inequality related to health in France on the national scale. Through an econometric analysis based on panel data from 2000 to 2004 at departmental level, I investigate the total mortality rate in relation to socioeconomic status and air pollution. The concentration level of NO2 , O3 and PM10 are estimated by spatial interpolation from local observations by a network of monitoring stations. I find a positive and significant relationship between NO2and the mortality rate, at mean levels below the current standard, with a greater relative risk for women. Moreover I observe disparities in health through income among French departments. These results not only confirm the existence of a relationship between current air pollution levels and mortality but also raise questions about environmental policy implications in France.
1.1. Introduction
Many pollutants are declining throughout the industrialized world. However, ex- posure to air pollution, even at the levels commonly reached nowadays in European countries, still leads to adverse health effects. In this context, there has been in- creasing global concern over the public health impacts attributed to environmental pollution.
Multilevel modelling has been previously used to assess the negative correlation between pollution exposure and socioeconomics status, such as unemployment, ed- ucation, and the working class in Canada (Premji et al., 2007), ethnic group, and
income in England (McLeod et al., 2000) and in the US (Grineski et al., 2007), (Morello Frosch et al., 2002) where the concept of environmental justice has been the object of increasing attention. Viel et al. emphasize that towns with high pro- portions of immigrants tend to host more hazardous sites even when controlled for population size, income, degree of industrialization of the town, and region (Viel et al., 2011). In Germany, Schikowski et al. show the existence of social differences in respiratory health among the female population (Schikowski et al., 2008) and Bolte et al. acknowledge social inequality in perceived environmental exposure in relation to housing conditions (Bolte et al., 2010). Pearce et al. for New Zealand point out that industrial pollution is greater in wealthy places, whereas overall pol- lution affects poorer zones more (Pearce et al., 2010) .
Moreover, multiple models also estimate the relationship between health and pollu- tion, showing the impact of outdoor air pollution on the mortality rate in Austria, France and Switzerland (Kunzli et al., 2000), in England (Janke et al., 2009), on the allergic sensitization on primary schoolchildren in France (Maesano et al., 2007), on asthma (Wilhelm et al. 2009), or on cancer risks among schoolchildren in the US (Chay et al., 2003), (Morello Frosch et al., 2002). Finally, Finkelstein et al. point out that mean pollutant levels tend to be higher in lower income neighbourhoods in On- tario and both income and pollutant levels are associated with mortality differences (Finkelstein et al., 2003).
In addition, the literature is far to be silent about the relationship between health and socioeconomic status (SES). A number of SES measures have been proposed,
including income, wealth, education, labor force status, and race/ethnicity. For in- stance, some recent papers provide evidence of a positive association between income and health (Subramanian & Kawachi, 2000), (Gunasekara et al., 2011). Apart from income, education has also been considered a crucial component of SES affecting health (Grossman, 2000). In France, Cambois & Jusot (2010) study the link be- tween lifelong adverse experiences, health and SES. Lifelong adverse experiences is related to poor self-perceived health, diseases and activity limitations even control- ling for SES. Results from Stringhini et al. (2012) suggest that the social patterning of unhealthy behaviors differs between countries. They stress health behaviors are likely to only be major contributors of socioeconomic differences in health. Among others, Lindahl (2005) focus on mortality rates and find a positive causal relationship between income and health measure.
Moreover, I observe a growing epidemiologic literature about the effects of air pollution on health by gender. The most recent gender analysis from Clougherty shows that most studies for adults report stronger effects among women, particu- larly when using residential exposure assessment (Clougherty, 2010). The smaller size of the trachea has been argued to be a reason which makes women more sen- sitive to particulates inthe air (Marr, 2010). However, it remains unclear whether the observed difference is a result of gender-linked biological differences or gender differences in activity patterns.
The analysis offers several contributions to the existing literature. Most inter-
national empirical economic studies estimate either the relationship between health and pollution or the correlation between pollution exposure and socioeconomic sta- tus. I aim to gather both literatures to assess the impact of air pollution on health according to social status. Few European studies investigated the effect modification of socio-economic factors on the association between air pollution and health and much is yet to be understood (Deguen & Zmirou, 2012). European policy-makers have in fact only recently acknowledged the notions of environmental justice and environmental inequalities, which have been part of the US policy arsenal for almost two decades (Laurent, 2011). To my knowledge, environmental factors affecting health, such as exposure to atmospheric air pollution have not been yet studied in France on a national scale in the context of social inequalities. Laurent et al. em- phasize the importance of continuing to investigate this topic due to the tendency for greater effects to be observed among the more deprived (Laurent et al., 2007).
Whereas the french literature only looks at high level of pollution, I am studying am- bient air pollution; the dataset presents low level of pollution concentration, below the actual threshold fixed by the public authorities at which health can be harmed (Pascal et al., 2009). Instead of looking at one geographical area, I examine re- cent relationships between pollution and health for the entire country using a panel dataset. I also account for unobserved confounders using fixed effects clustering at the regional level not to suffer from potential omitted variable bias. Most of the studies on this topic use times series or a cross sectional cohort (Janke et al., 2009).
Times series exploit short-term variation to identify pollutant effects which elimi-
nates the effects of lifestyle factors such as smoking, exercise and diet, because these factors do not change on the short run. The cohort studies may also suffer from omitted variables bias, as the cities or zip codes which are compared, may differ from each other in important ways other than just their levels of pollution. Some recent studies use a exogeneous event to cope with omitted variables biais (Chay
& Greenstone, 2003) (Moretti & Neidell, 2011), (Currie & Walker, 2011). For in- stance, Chay and Greenstone use a sudden recession as an instrument to identify the effect of a medium-term reduction of pollution on infant mortality (Chay & Green- stone, 2003). Finally, I use a model which takes into account spatial autocorrelation.
This paper investigates the relationship between ambient air pollutant concen- trations, social class, and population mortality on the departmental scale in France
1. It is part of new research on environmental justice, and provides an overview of the distribution of environmental risks. To identify the social distribution of air pollution, the study compared the social characteristics (income, unemployment) and the concentration of air pollution among French local authorities for different level of poverty. In this context, we first may wonder whether poor areas are also the ones with low socioeconomic level. Poor people may be more likely to live where pollution may be higher, next to industrial area (Mohai et al., 2009). Due to bud- get constraint, the unemployed people are also less likely to move from one area to another to avoid pollution. Secondly, we also ask oneself if a change in pollution
1. Department corresponds to a local authority below the regional level.
benefits, in term of health, even more to high socioeconomic than low levels areas.
When it comes to poor local authorities, is the health effect of an increase in air pollution twice over? The main purpose is to figure out if inequalities tend to mount up within French local authorities.
I find a positive and significant relationship between NO2 and the mortality rate, at mean levels below the current standard, with a greater relative risk for women. I show higher is the income level for French department, lower is the level of mortality rate. However, health disparities appear to be more related to socioeconomic factors than differences in sensitivity to pollution.
1.2. Medical perspective
The L.A.U.R.E (Law on Air and Rational Use of Energy) and the different Euro- pean directives give priority to monitoring common air pollutants with a direct effect on health, such as Nitrogen Dioxide (NO2), Nitrogen Oxides (NO), Ozone (O3) and Particles (PM10). In consequence, I consider these pollutants in this paper. The contamination of the atmosphere by pollutants at the local and regional level is the result of three processes: emission, transmission, and air pollution concentration.
Pollutants are first released at the source with gases and particles which are put into the air. The pollutants emitted are then dispersed, or sometimes they can be chem- ically transformed in the atmosphere, creating new, secondary pollutants. Having combined with air and become diluted, they create a concentration of toxic levels of
chemicals in the air, and these atmospheric pollutants are finally inhaled by humans, animals and plants.
First, Particulate Matter (PM) is made up of a number of components, including acids, organic chemicals, metals, and soil or dust particles. The size of particles is directly linked to their effect on health: PM10 (aerodynamic diameter less than 10 µm); PM2.5 (aerodynamic diameter less than 2,5 µm) are the particles that gen- erally pass through the throat and nose and enter the lungs. The PM2.5 particles are the most dangerous. The effects of PM on health occur at levels of exposure currently being experienced by most urban and rural populations in both developed and developing countries. Chronic exposure to particles contributes to the risk of developing cardiovascular and respiratory diseases, as well as of lung cancer (WHO).
Once inhaled, these particles can affect the heart and lungs and cause serious health effects. Not only have many European projects found a link between particles and mortality or morbidity (Peng et al., 2004) (Touloumi et al., 1997), but so have recent epidemiologic studies (Schikowski et al., 2008), (Janke et al., 2009), (Maesano et al., 2007).
Nitrogen Oxides (NOx ) is the main indicator of transportation and stationary combustion sources, such as electric utility and industrial boilers contamination2.
2. The spatial distribution of NO2 is generally not homogeneous within individual metropolitan areas.
The primary reason for the observed heterogeneity in concentrations across an urban area is the sub- stantially higher concentrations of NO2 near sources, such as roads [Electric Power Research Institute 2009].
NOx forms when fuels are burned at high temperatures and includes various Nitro- gen compounds such as Nitrogen Dioxide (NO2) and Nitric Oxide (NO). Nitrogen Dioxide (NO2) and Nitric Oxide (NO) play crucial role in the atmospheric reactions by creating harmful particulate matter, ground-level Ozone, acid rain, and eutroph- ication of coastal waters. NO2 is produced by chemical transformation with NO and Ozone (NO + O3 = NO2 + O2). Not only particle filters but also the rise of Ozone in the atmosphere increase NO2 emissions (AFSSET). As a consequence, NOx is a powerful oxidizing gas, linked with a number of adverse effects on the respiratory system (Agency, 2011).
Ozone (O3) is an example of a secondary pollutant as it is formed when Hy- drocarbons (HC) and Nitrogen Oxides (NOx) combine in the presence of sunlight.
And excessive Ozone in the air can have a marked effect on human health. It can cause breathing problems, trigger asthma, reduce lung function and cause lung dis- eases (OMS ). Breathing ozone can trigger a variety of health problems including chest pain, coughing, throat irritation, and congestion (Agency, 2011). Recent epi- demiologic studies emphasize the relationship between Ozone and the mortality rate (Janke et al., 2009) and asthma exacerbation (Currie & Neidell, 2004), (Laurent et al., 2007), (Wilhelm et al., 2009).
1.3. Presentation of the dataset
I use data on the concentration of pollutants and mortality rates available at a local level for all the whole of France.
Detailed data on atmospheric pollution come from the information system of the air quality measure (BDQA) used by the French Environment and Energy Manage- ment Agency (ADEME). ADEME gathers information coming from the 38 associ- ations (AASQA) within the ATMO federation which monitor air quality. A large number of monitoring stations make up the federation. The French nomenclature identifies seven classes of stations, consistent with the various classifications defined at the European level : roadside, urban, industrial, near city background, national rural, regional rural, specific observations numbering 84, 286, 119, 138, 10, 62, and 13 respectively. Most of the monitoring stations are placed where the density of pop- ulation is significant, apart from national rural monitoring stations. The measure taken into consideration in the study is the annual mean of concentration for pollu- tants within a civil year (1st January to 31st December) calculated by each AASQA for each captor and measured in micrograms per cubic meter of air. In principle the more disaggregated data is more desirable to cope with ecological inferences, but the health authority estimates are based on surveys with relatively small samples and are therefore less reliable. However, aggregate data may offer valuable clues about individual behavior. I divide the dataset in subsamples as an attempt to deal with the problems of confounding and aggregation bias. This annual mean is calculated by the ASQAA from the hourly mean for each monitoring station. This
unit of concentration is mostly used to monitor outdoor air quality. Air pollutant concentrations do not necessarily produce accurate predictions of exposure levels.
People may be resident in one area, but work in another. Nevertheless, the geo- graphical level used in this article reduces the bias related to population mobility.
The department surface represents an average of 570 000 hectares and we know from INSEE data that the average distance between the place of residence and the place of work is nearly 20km, so the accuracy of the exposure levels seems reasonable.
For spatial interpolation between monitoring stations, I use a geostatistical method that takes into account spatial dependence. This method does not necessarly reduce the amount of measurement error in the variable. The extent of measurement error is going to be greater for those departments with few monitoring stations where the population is more dispersed or lower. Lower or higher levels of the dependent vari- able within departments also induce measurement error. For example, more rural areas tend to be more agriculturally based and this may have an impact on mor- tality rates. Nevertheless, measurement error, even if not systematic, can induce attenuation bias.
Following Currie & Neidell (Currie & Neidell, 2004), I assign annual pollutant concentrations to the 95 French departments. Using the geographical coordinates of the census blocks with the highest population density of a local authority, I calculate the distance between the census blocks with the highest population density and all
monitoring stations as it is explained in Appendix 1. First, I calculate the centroid of each local authority. I then measure the distance between the monitoring station and the center of the local authority. This distance corresponds to the weight at- tributed to a monitoring station, using the inverse of the distance to the center of the local authority. In order to assess the accuracy of our measure, I compare the actual level of pollution at each monitor location with the level of pollution that I would assign using the method previously described. The correlations between the actual and predicted levels of pollution are quite high for O3, NO2 and PM10 (0.6, 0.85 and 0.7 respectively) suggesting that the measure is quite accurate.
Table 1.1: Summary statistics
Variable Mean Std. Dev. Min. Max. N
Pollutant variables
NO2µg/m3 31.714 11.375 12 74.046 220
O µg/m3 53.029 15.39 30.601 99.653 220
PM10µg/m3 21.461 5.508 8.818 57.384 220
Mortality rates
Overall mortality rate 819.759 64.615 620 1000 220
Ms (%) 626.536 47.623 499 756 220
Mr (%) 1092.595 96.551 792 1393 220
Socioeconomics variables
Income (%) 15303.625 2703.677 11011.659 27079.313 220
Un (%) 8.436 2.033 4.575 14.625 220
Education (%) 15.678 5.045 10.241 37.481 220
Poverty gap 0.5 0.501 0 1 220
Weather variables
Sun (hours) 1979.783 362.492 1367.4 2962.3 205
Pr (mm) 2.211 0.571 0.855 3.865 214
Wind (km/hour) 100.212 13.056 68.400 147.6 208
Frost (days) 41.61 22.186 4 114 213
Demographics variables
Sm (%) 1218.513 274.182 483.5 2298.7 220
Pop 771461.595 539704.805 123561 2561038 220
Industry (%) 16.175 4.778 5.93 24.476 220
PPHB (%) 136.174 37.84 71.696 268.82 220
Alcohol (%) 262.614 223.659 60 1594 220
Accident (%) 12.055 4.107 3 22 220
Atmo Index
Atmo index 8 to 10 3.524 5.532 0 28 220
The top panel of Table 1.1 presents descriptive statistics for pollution data. NO,
NO2 and PM10are positively correlated with correlation coefficients between 0.5 and 0.8 as we can see in Table 1.2. They are negatively correlated with O3 which may be due to the fact that Ozone is rapidly destroyed to form NO2 within cities the correlation between both NO2 and NO is high (0.85), so that I choose to keep NO2 as an explanatory variable and drop NO to prevent autocorrelation. Moreover, I do not include observations for SO2 and CO, as few monitoring stations measure these pollutants.
Table 1.2: pairwise correlation coefficients with significance level
NO2 PM10 O NO Atmo index(8-10) Temperature
NO2 1.0000
PM10 0.6086 1.0000
(0.0000)
O -0.3238 0.0301 1.0000
(0.0000) ( 0.6566)
NO 0.9644 0.5921 -0.3265 1.0000
(0.0000) (0.0000) (0.0000)
Atmo index(8-10) 0.2663 0.5121 0.3833 0.1473 1.0000
( 0.0034 ) (0.0000) (0.0000) (0.1099)
Temperature 0.1818 0.3079 0.4762 0.1447 0.2593 1.0000
(0.0077) (0.0000) (0.0000) (0.0344 ) (0.0044)
Other pollutants are also likely to be associated with differences in mortality, but data were unavailable to perform intra urban interpolations for these pollu- tants. Note that the local authorities with missing air pollution measures are all less populated areas. It is important to stress that air pollutant concentrations used to be below the limit value fixed by European and national institutions above which health can be harmed. In France, the threshold for NO2, fixed by the European act 2002-13 related to air quality, is 200 µg/m3 over 24 hours to protect human health.
For a long term exposition (over one year) the regulated levels is 40µg/m3 with re- spect to World health organization. The maximum concentration of NO2 presented in the dataset is over this threshold. However, the annual mean for NO2 is below
the regulated level. The annual mean is the measure I use in my estimations. It corresponds to ’very good’ air quality according to the ATMO index presented in Table 9. Moreover, the average concentration from the measure is lower than the level used by studies in the United States and even lower than in England (Janke et al., 2009), where the level at which it is considered to harm health is already quite low. However, the E.R.P.U.R.S project in France shows that NO2 and PM10 have a negative impact on health, even at low air concentrations, considering hos- pitalization numbers as the explicative variable (Campagna et al., 2003). Pascal &
al. in France obtain similar results, considering also different mortality rates in nine polluted cities (Pascal et al., 2009) .
The second panel of Table 1.1 presents non incidental mortality rates. I consider a period of 5 years (2000-2004). The year corresponds to mid-year of the triennial period used. A moving average makes it possible to "smooth" a series of values expressed according to time. It is used to eliminate the least significant fluctuations.
Mortality rate is a moving average of order 3. Data on mortality are available from 1980 to 2004 whereas data on pollution only exist from 1985 to 2005 with very few values before 2000. A large range of pollutants are responsible for outdoor air pollution, so that it is difficult to assign them to a specific health effect. This is why I use an non incidental mortality rate. We do not include in this paper specific causes of mortality, due to the weak variability of these data in France for the 2000
- 2004 period which does not allow any estimations. For instance, transport release N02 such that a high level of pollution may be observed next to roads where road accidents occur. As I wish to extract the only effect of pollution on mortality, I work on non incidental rate. The data on health come from the National Federation of Regional Health Observatories (ORS). I use age-standardized rates to control for different age structures across departments3. The standard deviation is quite high, showing that the data are spread out over a large range of values. The degree of dispersion (spread) and skewness in the data are presented graphically in Figure 1.1.
Figure 1.1: The yearly distribution of all causes mortality rates for the 2000-2004 period, in all departments.
3. This age-standardized rate is calculated as follows: P19
i=1P iT i. Pi represents the share of age group for the population of reference and Ti represents the specific rate of mortality observed within a department for the age group i.
The third panel shows the socioeconomic variables: Income, education, poverty gap and unemployment based on the 2007 census of INSEE and the French Ministry of Labour (DARES). Definitions of the variables are given in Table 1.3. Note that data about ethnicity or race do not exist in France. The French Institute of Statistics does not collect data about language, religion, or ethnicity on the principle of the secular and unitary nature of the French Republic.
Table1.3:Definitionofvariables VariableDefinitionSources MTotalmortalityrate,agestandardizedrates2000-2004calculatedusingdataonregistered deathsfromINSERM,CEPIDcandINSEE.Theyearcorrespondstomid-yearofthetrienal periodused.Unit:per100000people Nationalfederationofregionalhealth observatories(ORS) MrTotalmalemortalityrate,agestandardizedrates2000-2004calculatedusingdataonregis- tereddeathsfromINSERM,CEPIDcandINSEE.Theyearcorrespondstomid-yearofthe trienalperiodused.Unit:per100000people
Nationalfederationofregionalhealth observatories(ORS) MsTotalfemalemortalityrate,agestandardizedrates2000-2004calculatedusingdataonreg- istereddeathsfromINSERM,CEPIDcandINSEE.Theyearcorrespondstomid-yearofthe trienalperiodused.Unit:per100000people
Nationalfederationofregionalhealth observatories(ORS) NO2,PM10,O3AnnualmeanofNO2,PM10,O3concentration(µg/m3)respectively2000-2004FrenchEnvironmentandEnergyMan- agementAgency(ADEME) PrHighprecipitationtotals2000-2004MétéoFrance SunAnnualcumulationofinsolationinhours2000-2004MétéoFrance FrostAnnualnumberoffrostdaysindays2000-2004MétéoFrance WindAnnualinstantaneaousmaximumwindinkm/hour2000-2004MétéoFrance SmNumberofcigarettessoldfor1000residents2000-2004FrenchMonitoringCentreforDrugs andDrugAddictions(OFTD) AlcoholAnnualnumberofdeathsrelatedtoalcohol.Itincludeslivercirrhosis,alcoholicpsychosis andalcoholism,canceroftheupperaero-digestivetractNationalfederationofregionalhealth observatories(ORS) AccidentRoadaccidentrate,agestandardizedrates2000-2004calculatedusingdataonregistered roadaccident.Unit:per100000people.Nationalfederationofregionalhealth observatories(ORS) PPHBNumberofpeopleper1hospitalbed2000-2004Nationalfederationofregionalhealth observatories(ORS) IndustryShareofindustryinthetotalvalueaddedofadepartment(in%).FrenchNationalInstituteforStatistics (INSEE),census2005 EducationPopulationfrom15years(withoutstudents)withminimumBAC+2dividedbypopulation withindepartmentin2006FrenchNationalInstituteforStatistics (INSEE),census2007 UnTheunemploymentrateisthepercentageofunemployedpeopleinthelabourforce(occupied labourforce+theunemployed)2000-2004.FrenchMinistryofLabour(DARES) IncomeIncomeisdefinedasthenettaxableincomedividedbythenumberoftaxhouseholdswithin adepartmentFrenchNationalInstituteforStatistics (INSEE) Theintensityof poverty(orpovertygap)isanindicatorusedtoassesstheextenttowhichthestandardoflivingof thepoorpopulationisunderthepovertyline.Itiscalculatedformallyasfollows:(poverty threshold-medianstandardoflivingofthepoorpopulation)/povertythreshold.
FrenchNationalInstituteforStatistics (INSEE),census2005
The following panel describes the control variables. Data on weather come from Meteo France through the French Institute of the Environment (IFEN). Smoking rate fell by 35% between 2000 and 2004, probably due to the "Loi Évin" of 1991 and the tax increase (INSEE ). Road accident rate fells 29% according to the data from the National Federation of Regional Health Observatories (ORS). I also collect from the ORS the number of people per hospital bed to measure the health care system and the availability of medical care resources in a particular department from 2000 to 2004. I add the share of industry to control for industrialization,4 as a time invariant variable for each department based on the 2005 census of the French Institute of Statistics (INSEE).
Finally, the last row of descriptive statistics corresponds to an air pollution index.
To capture peaks of pollution, I use the ATMO index calculated by the AASQA.
The Atmo outlook varies daily according to air quality using a scale of 1-10 (1 = very good air quality, 10 = very bad air quality). This index takes into consideration the concentration of four subindexes characterizing Nitrogen Dioxide (NO2), Sulphur Dioxide (SO2), Particles in suspension (PS) and Ozone (O3). It considers pollution measured only by urban and industrial monitoring stations for main agglomerations for a period from 2000 to 2003. After 2003, the construction of the index was changed, so that I cannot consider it for 2004. I retain 41 agglomerations and I associate each one with a department. I construct a yearly variable summing up
4. French data about industrialization and GDP are not precise enough to take into account time fluc- tuations among departments from 2000 to 2004. pourcentage of industry added-value over the total-added value for each department is available only every five years (INSEE).
the number of days above indices 8, 9 and 10, which corresponds to poor air quality according to the definition of the Atmo index (Table 1.4).
Peaks of pollution are correlated positively with ambient air pollution which gives more credence to the measure. This index variable is positively correlated with the previous measure of NO2, NO, PM10 and O3. However, further in the estimation, I prefer to use real concentrations of pollution instead of indices. In fact, few days correspond to peaks of pollution, and fixing a threshold below which pollution does not have any impact is highly arguable. Pollution does indeed fluctuate, a low level can be active and the level perceived as toxic is variable, even among the healthy population. Within a population, some people are more sensitive than others and will suffer from atmospheric pollution even at really low levels ; levels below the actual threshold fixed by the public authorities. I aim to test this intuition.
Table 1.4: The Atmo index
P M10scale N O2 scale O3 scale
Index scale Subindexes Average of mean daily concentrations in µg/m3 Average of the hourly maxima in µg/m3
Very good 1 0 - 9 0 - 39 0 - 29
Very good 2 10-19 40 - 79 30 - 54
Good 3 20 - 29 80 - 119 55 - 79
Good 4 30 - 39 120 - 159 80 - 104
Moderate 5 40 - 49 160 - 199 105 - 129
Poor 6 50 - 64 200 - 249 130 - 149
Poor 7 65 - 79 250 - 299 150 - 179
Bad 8 80 - 99 300 - 399 180 - 209
Bad 9 100 - 124 400 - 499 210 - 239
Very bad 10 125 and more 500 and more 240 and more
1.4. Model and Econometrics
1.4.1. Specification
The focus of this study is the relationship between average pollution, socioe- conomic status, and mortality. the unit of analysis is the department, which is the main administrative unit below the national regional level. The department of France are French administrative divisions. The departments form one of the three levels of local government, together with the 22 metropolitan and 5 overseas regions above them. There are 95 departments in France with an average population of 620 000 people, ranging from over 70 000 to over two million. Departments are grouped within 22 metropolitan areas known as regions5.
In the analysis, I start by estimating a standard model with the non incidental mortality rate as the explicative variable without considerations of environmental quality. After doing preliminary regressions for various functional forms and follow- ing the results from an overall normality test based on skewness and on kurtosis for each of them, I estimate an equation of the following form to ensure that errors are normally distributed ε ∼ N (0, σ2) 6 :
Xitk = αk+ Socioeconomicitβk+ Demographicsitηk+ Zitφk+ εkit (1.1)
5. Due to missing data, we remove departments from the analysis in order to consider a balanced panel.
We end up with 41 departments
6. There is no evidence that the log transform is the best fit for mortality time trends (Bishai & Opuni, 2009). Moreover, given the size of the department, the effect of outliers may not be a problem here.
where i indexes the local authority, t indexes the year, k the kind of mortality rate. Xitk is a vector of all causes mortality rates (overall mortality rate, male and female mortality rate). Socioeconomic variables and particularly the unemployment rate and income are included as the main explanatory variables. Due to multi- collinearity issues, I am not including both the average income and the education variable. The squared correlation between education and the average revenue is above 0.8.
The vector Demographicsit includes several variables. First, it accounts for lifestyle, which refers to the regular activities and habits a person has that could have an ef- fect on his or her health. I include the smoking rate variable as a proxy for lifestyle.
The number of people per hospital bed P P HBit in each department is included as a proxy to measure the health care system and the availability of medical care re- sources in a particular department. I also include the % of industry added-value over the total added-value Industryi for each department as a time-invariant variable. I also take weather patterns into consideration at department level Zit as a control for average pollution levels. I consider the annual mean of precipitation to capture the effect of very wet years, the maximum wind speed, the number of frost days, and the annual cumul of sunlight as a time varying control. Some studies contend that mostly long-term (i.e., monthly and annual) fluctuations in temperature affect mortality (Martens, 1998). Besides, wind speed measurements are important for air quality monitoring. The higher the wind speed, the lower the pollutant concentra- tion. Wind dilutes pollutants and rapidly disperses them throughout the immediate
area. εit is the error term.
Independent variables have explained most differences between departments and years, but there is probably some unmodeled heterogeneity. Thus, the next step in performing a multilevel analysis is to decide whether the explanatory variables considered in the analysis have fixed or random effects. The Hausman test considers the null hypothesis that the coefficients estimated by the efficient random effects es- timator are the same as the ones estimated by the consistent fixed effects estimator.
By running this test, the fixed effect model appears to be the most efficient one. In fact, I think of each department as having its own systematic baseline. I also calcu- late the robust variance estimator, in order to prevent the heteroskedasticity that I found by running Breush-Pagan test: this test checks if squared errors are explained by explanatory variables. the estimation will also take into account autocorrelation, because the Wooldridge test shows that disturbances exhibit autocorrelation, with the values in a given period depending on the values of the same series in previous periods. To address the possibility that omitted variables account for some of the heterogeneity among French departments, an error component model is estimated:
εit = ci+ δt+ uit (1.2)
ci and δt are residual differences where ci is a department effect which accounts for differences across departments that are time-invariant (e.g lifestyle differences that we cannot take into account ), δt is a year effect which controls for factors that
vary uniformly across departments over time, and uit is the remaining error term 7. It is also likely that a population’s health affects unemployment via productivity, education and other factors. This potential simultaneity can be a source of endo- geneity, making standard estimators inconsistent I need to test this hypothesis, so I consider the lag of the endogenous variable, unemployment, as an instrument. The F-test on the excluded instruments in the first stage regression confirms the validity of the instrument. To avoid the weak instruments pathology, we look at the F-test on the excluded instruments in the first stage regression and check whether the test statistic is greater than 10 (F(1,192) = 28.91). Then, the Hausman test rejects the endogeneity of the model (P=0.810).
This paper is also concerned with spatial correlations which would bias the results or introduce inefficiency. If the observations are spatially clustered, the estimates obtained will be biased or inefficiency will be introduced. In fact, the mortality rate in one region could be related to that in another. 8. As a consequence, I will cal- culate the Driscoll and Kraay non-parametric adjustment of standard errors model allowing for both space and time adjustments.
In the second model, the mortality rate is expressed as a function of environmen-
7. The Ramsey test confirms the robustness of the specification.
8. The Moran Index of spatial contiguity rejects the null hypothesis that there is no spatial clustering of the value in the raw mortality data. First tail test: I=0.266 at a 1 pourcent probability
tal variables added to the previous variables. I will estimate the model :
Xitk = λk+ Pitθk+ Socioeconomicitψk+ Demographicsitσk+ Zitφk+ εkit (1.3)
Pit is a vector of air pollutant concentrations for O3, NO2 and PM10. In this model, the main coefficient of interest is θ representing the mean parameter estimates for all 95 departments for explanatory variables Pit. It also represents the effect of air quality on health outcomes. I will again use the fixed effect estimator to control for heterogeneity between departments, with and without the Driscoll and Kraay standard errors model. Besides, the endogeneity problem has to be discussed in this context. I include fixed effects and some controls to address the problem of unob- servable variables. However; there may be time-varying unobservable variables, not common to all regions and not captured by the dummies, which could bias the esti- mates. One may argue that European unemployment fluctuates around a very low level (Blanchard, 1986) making the previous Hausman test really weak. An associ- ation between the business cycle and mortality could, for instance, be driving the result (Chay & Greenstone, 2003). However, the French Statistical Institute does have access to the yearly business cycle data for each department.Another endo- geneity bias could be that people may move in response to pollution levels. People who care more about health, and hence live a healthier life style, may move to less polluted areas, introducing an upper bias in the estimate of pollutants. However, migrations in France between departments are essentially in border areas and are mainly due to preferences for urbanization (INSEE). .
Finally, I estimate a model dividing my sample with those above and below the median of poverty gap. To do so, I create a dummy with respect to the intensity of poverty. 50 % of departments are below and 50% are above this median. I want to study whether the impact of pollution on health is greater when I consider poorer population. Figure 1.3 shows a potential positive relationship between NO2 and the intensity of poverty.
Figure 1.2: Correlation between poverty gap and NO2
Departments with a high pollution level seems to be the one with a low socioe- conomic level. Besides, I ask oneself if a change in pollution benefits, in term of health, even more to high socioeconomic than low levels areas. The main purpose is to figure out if inequalities tend to mount up within French department. People with low incomes may be disproportionately vulnerable as well as disproportionately
exposed. Is the health damages of an increase in air pollution bigger in poorer area compared to its counterparts? I show in the next section that socioeconomic factors, in particular the unemployment rate, greatly interfer when studying the impact of N O2 on mortality rate at different level of poverty.
1.4.2. Results
1.4.2.1. Impact of environment quality on health
I start by examining a standard model of mortality rate without consideration of environmental quality. I then add NO2, O3 and PM10 to the specification to see if considering pollutant variables improves the global fit of the model. To capture the department effect and the spatial autocorrelation, both fixed effects clustered at the regional level and the Driscoll and Kraay standard errors with fixed effect are estimated. Approximately seventy percent of the variation in the response variable may be attributed to explanatory variables.
I first estimate the standard model with the OLS, trying to test the most com- plete model, and I observe in the first column from Table 1.5 that all the coefficients of the determinants of mortality are significant. However, I also use the within esti- mator as I assume that the unobserved factors fit between departments determine both mortality rates and explanatory variables. I observe a loss of significativity for some coefficients which may be due to the correlation between department-specific effect and both explanatory variables. The Fixed effect imposes time-independent
Table 1.5: Standard model of non incidental mortality rate
OLS FE D-K S.E
Income -0.0165*** -0.0641*** -0.0641**
(0.00253) (0.00596) (0.0141)
Un 8.217* 17.01** 17.01***
(4.435) (7.432) (2.593)
Pr -16.83 3.755 3.755
(9.914) (5.898) (5.744)
Sun -0.0793*** 0.0764*** 0.0764**
(0.0178) (0.0174) (0.0206)
Wind 0.786*** 0.271** 0.271
(0.267) (0.120) (0.199)
Frost 0.728** -0.966*** -0.966*
(0.290) (0.209) (0.425)
Sm 0.0619** -0.105*** -0.105*
(0.0231) (0.0247) (0.0412) Industry
PPHB 0.0317 1.690 1.690
(0.237) (2.411) (0.876)
Alcohol 0.0840 0.174* 0.174**
(0.0508) (0.0901) (0.0613)
department FE x x
Observations 203 203 203
R-squared 0.598 0.753
a
a. Notes: This table presents the standard model of non incidental mortality rate with its main determinants. All regressions are estimated with standard errors clustered at the regional level. Robust standard errors in parentheses. Statistical significance is denoted by: *** p<0.01,
** p<0.05, * p<0.1
effects for each entity that are possibly correlated with the regressors, which is why Industryi, a time invariant variable, is not taken into account. The last column shows the Driscoll and Kraay standard errors model which takes spatial autocor- relation into account, with fixed effects. I observe that income impacts negatively mortality rate in every regressions at a 1% level of significativity. I am in line with the literature saying that income is a significant determinant of health. Et- tner finds that increases in income significantly improve mental and physical health (Ettner, 1996). Inadequate education and living conditions ranging from low in- come to the unhealthy characteristics of neighborhoods and communities can harm
health through complex pathways. Health disparities by income is partly explained by disparities in medical care. French departments with a high level of income are more likely to have a low mortality rate than department with a lower income. As a consequence, a shortsighted political focus on reducing spending in education, child care, jobs, community and economic revitalization, housing, transportation, could actually increase medical costs by magnifying disease burden and widening health disparities.
I then study the relationship between NO2, O3, PM10and mortality rates in both a single pollutant model (Table 1.6) and in a multi-pollutant one (Table 1.7). The multi-pollutant model allows coefficients to be examined at the same time, so as to not overestimate the impact of one pollutant.
As shown in the single pollutant model, coefficients for PM10 are significant with the fixed effect model clustered at the regional level and the Driscoll and Kraay estimation. However, coefficients are not significantly different from zero when I consider a multiple pollutant model in both specifications. The variation I have in the dataset may not be sufficient to obtain significant results for PM10 even though the single pollutant model shows a significant impact of PM10 on mortality rates.
Mortality rates do not vary too much within region over time. moreover, we only have access to the variation within departments over a few years, from 2000 to 2004.
This may be very little variation, perhaps some of it due to measurement error which would bias coefficients towards zero. It also may be explained by the interaction be-
Table 1.6: A simple pollutant model of mortality
(1) (2) (3) (4) (5) (6)
VARIABLES FE D-K S.E FE D-K S.E FE D-K S.E
NO2 0.438* 0.438*
(0.237) (0.182)
O 0.220 0.220
(0.377) (0.195)
PM10 0.193 0.193*
(0.263) (0.0881)
Income -0.0645*** -0.0645** -0.0646*** -0.0646*** -0.0649*** -0.0649***
(0.00568) (0.0142) (0.00543) (0.0134) (0.00560) (0.0134)
department FE x x x x x x
Weather controls x x x x x x
Demographic controls x x x x x x
Socioeconomic controls x x x x x x
Observations 203 203 203 203 203 203
Adjusted R-squared 0.756 0.754 0.754
a
a. Notes: This table presents the impact of N O2, O3and P M10on non incidental mortality rate.
All regressions are estimated using fixed effect with standard errors clustered at the regional level or with Driscoll and Kraay standard errors. I include in all estimations a vector of weather pattern with wind, sun, precipitations and frost; a vector of socioeconomic variables including unemployment rate and income; and a vector of demographics including the level of indutrialization, people per hospital bed and the smoking rate. Robust standard errors in parentheses. Statistical significance is denoted by: *** p<0.01, ** p<0.05, * p<0.1
tween O3 and NOx which may biased the result obtained for PM10. This result is in line with Chay et al. who examine the effect of particulate matter on adult mortality in the US during the 1970s. They find no impact of this source of pollution on adult mortality (Chay et al., 2003). In contrast, this result is opposed to the French study by the Sanitary Health Institute which found a positive effect of PM10 on mortality in a panel of nine different French cities (Pascal et al., 2009). However, this study does not precise the type of estimator used. Furthermore, Pascal et al. do not take into account the influence of lifestyle or socioeconomic factors on health as their model strictly includes weather data whereas the robustness of the model is not verified if I take socioeconomic factors out. Finally, the average concentration from the measure is probably lower than the level used by the Sanitary Health Institute
